Wave, Stripe Kai skills gating, AgentCore Runtime V2 snapshots
Systems & planning
Amplitude Wave ships product improvements in 1–2 days…
TL;DR: Wave, Amplitude’s proactive product agent built primarily on OpenAI models, identifies what to build, helps teams ship changes with coding agents, and measures impact.
- Wave is a proactive product agent built primarily on OpenAI models.
- It identifies what to build and helps teams ship changes with coding agents.
- It measures the impact of the changes.
Source: https://openai.com/business/customer-stories/amplitude
Stripe Kai reached over 5
TL;DR: Stripe’s Kai connects employees to tools and data with Deep Agents, reaching 5,000+ users in about four weeks.
- Deep Agents supplies tool-calling, middleware, streaming, and state management; Stripe layers on its security, infrastructure, configurable agents, and UI.
- An S3-backed virtual filesystem syncs files into and out of sandboxes, preserving them across turns; the sandbox supports Python analytics and processing arbitrary file formats.
- Skills gate tool loading across 500+ MCP tools and 1,000+ skills; quality degrades with more than 150 skills combined with the system prompt, under a 1024 character frontmatter limit.
Why it matters: Kai’s tool catalog is too large to place in-context at once, so it uses skills to decide which tools to load on demand. The report’s “>150 skills + system prompt” degradation indicates a practical ceiling where tool/skill routing becomes necessary just to preserve baseline model quality.
Signals: Stripe is developing hybrid skill selection: pure LLM selection works at current scale, but the team expects to need a RAG or classifier layer to prefilter the growing catalog.
Source: https://www.langchain.com/blog/how-stripe-built-their-knowledge-ai-platform-on-deep-agents
Strands Box adds Linux support for agent sandboxes
TL;DR: Strands Box enforces file, network, credential, and Dogwood policy controls for agents running in Linux environments.
- Linux support is in preview for ARM64 hosts and works on Kubernetes, AgentCore Runtime, and Linux hosts across cloud providers or your own infrastructure.
- Box’s egress gateway injects credentials into approved requests; agents receive placeholders and cannot read the underlying secrets.
- Dogwood temporal rules can check prior actions and results; replacement requests require a matching HTTP 200 eligibility response within 300 seconds.
- On EKS, three deployment failures within five minutes trigger a ten-minute block on retries, while status checks remain allowed.
Signals: Linux support is in preview and currently runs on ARM64 hosts.
Source: https://strandsagents.com/blog/strands-box-linux-sandbox/
LangChain GTM Agent
TL;DR: LangChain’s GTM Agent generates deal-stage-specific artifacts; 47% of deals with sent artifacts advanced versus 16% of matched deals.
- The agent uses stage rules, templates, brand guidance, and Salesforce, usage, call, and Drive data; LangChain focused more on business context than prompt tuning.
- Slack requests or meeting prep trigger it to read Salesforce and Drive, ask for missing inputs, fill templates in a sandbox, render and review files, then hand them off.
- A stopping rule ships after one review finds no real defect, cutting one-pager generation time from 18 to 7 minutes.
- Over 12 weeks, it generated 1,540 artifacts; 46% were followed by a customer call within 14 days, and sent artifacts saved reps 337 hours.
Why it matters: The deployment ties artifact generation to deal stages, shared account context, and rep review rather than treating document creation as a standalone prompt. Its evaluation tracks utilization and deal outcomes alongside generation volume, giving teams signals about whether generated work is reaching customers.
Signals: The next version aims to connect generated artifacts to what enters a deal and why other artifacts go unused; an upcoming post will cover outcome measurement.
Source: https://www.langchain.com/blog/gtm-agent-high-quality-sales-artifacts
Strands Box brings Dogwood semantic policy to local…
TL;DR: Strands Box combines kernel-enforced isolation and Dogwood policies to control agent actions across tools, code, filesystems, and networks.
- Dogwood can require a successful
npm testsince the lastgit addand within 15 minutes beforegit push, or cap AWS MCP calls at 60/hour. - Box works with any model, agent framework, or harness, including agents that invoke tools directly, via MCP, or through shell or Python code.
- Its trusted computing base includes interpreters and protocol interceptors; for multi-tenant cloud use, the authors recommend a per-session MicroVM.
Why it matters: For multi-tenant cloud deployments, Box’s larger trusted computing base creates more opportunity for bugs that could bypass policy. The authors recommend pairing it with dedicated per-session MicroVMs to provide a strong boundary between agent sessions.
Signals: The team identifies making policies easier to write, maintain, and understand as a major area of investment over the coming months.
Source: https://strandsagents.com/blog/strands-box-the-big-picture/
- Dogwood can require a successful
Amazon Bedrock AgentCore Runtime V2 snapshots…
TL;DR: AgentCore Runtime V2 snapshots initialized environments to reduce new-session startup time; prewarming can hide startup for interactive users.
- V2 builds snapshots during create or update, which takes a few minutes; each new session restores memory and disk in a fresh microVM.
- Move imports, clients, and static data needed on the first request into init before app.run(); memory returns with the snapshot, but the file cache is cleared before the freeze.
- Build hooks warm memory before freezing and read files after a test restore; both calls share 120 seconds, and failure fails create or update without retry.
- A running session answers in about 200 ms; prewarm on app or chat engagement to hide startup, at a session cost for users who never prompt.
Why it matters: Runtime V2 makes session latency sensitive to what runs before the snapshot is taken: imports/clients/static data placed in initialization can be restored from memory, while any reliance on the file cache will not persist because it is cleared before freezing. Prewarming on user engagement can avoid “first prompt pays startup,” but you trade that latency for increased session utilization from users who never send a prompt.
Signals: Measure end-to-end session start before and after moving work into initialization: larger snapshots can offset the reduction in handler time.
Flint-interactive compiles semantic events into visual…
TL;DR: Flint-interactive lets agents specify chart interactions semantically while its compiler handles visual updates and connected views.
- Preset linked-brush highlights groups selected by brushing one view; viewport operators support panning or zooming.
- In chat, agents can generate an interaction_spec with a Flint chart spec, and users can pass selections as context to follow-up turns.
- Canvas event handlers can use semantic events to filter tables, update other charts, or change application state.
- Custom handlers map semantic events to update sequences, such as highlighting and labeling a clicked country without manipulating renderer-specific marks.
Signals: Interaction designers can turn bespoke interactions into reusable presets and propose them on GitHub.
Google Antigravity proposes rigidified macrocycles in an…
TL;DR: Google Antigravity’s cheminformatics skills curate compounds, analyze data, and propose rigidified macrocycles with synthesis routes in an hour.
- The model handles orchestration, database navigation, and reasoning; deterministic tools run calculations, grounded in RDKit and connected to ChEMBL and PubChem.
- The agent retrieved thousands of literature compounds, filtered buffers and solvents, linked compounds to reported data, and displayed results in an interactive viewer.
- Sparse data, inter-lab assay noise, and narrow dynamic ranges limit model precision; proposed designs remain hypotheses requiring synthesis and experimental validation.
Why it matters: Antigravity shows a pattern for agentic scientific workflows where the LLM orchestrates retrieval and reasoning while deterministic tooling (e.g., RDKit) does calculations and grounding against databases. For anyone building similar agents, the item also makes the evaluation boundary explicit: outputs are design hypotheses, and the cited data limitations mean you should plan for downstream experimental validation rather than treating proposals as confirmed improvements.
Signals: Proposed designs must be synthesized and validated experimentally.
Source: https://deepmind.google/blog/agentic-cheminformatics-with-google-antigravity/
Evaluation, grounding & uncertainty
ChatGPT Wallet
TL;DR: ChatGPT Wallet lets agents in ChatGPT Work, Codex, and dots buy with user approval and existing card protections.
- ChatGPT Wallet is intended for agents in ChatGPT Work, Codex, and dots.
- Agents can complete purchases with user approval.
- Purchases use existing card protections.
Why it matters: It provides an execution path for purchase actions where the transaction is gated by explicit user approval and relies on existing card protections, which reduces the blast radius when agents initiate payments inside ChatGPT Work, Codex, or dots.
BAIR CoRL 2026 schedule lists 33 papers and 10 orals and…
TL;DR: BAIR’s CoRL 2026 schedule lists 33 main-conference papers, four late-breaking posters, and 10 orals and spotlights.
- Monday’s Agentic Robotics late-breaking poster session runs 12:00–1:30 PM and lists “Scaling Online Exploration,” “Reconstruct, Practice, Go Real,” and “Flip Anything.”
- “Reconstruct, Practice, Go Real” is also scheduled for the Agentic Robotics workshop Thursday, November 12, from 4:15–5:55 PM.
- Wednesday’s 8:30–9:00 AM spotlight on optimization, agentic planning and evaluation includes “Playful Agentic Robot Learning” and GaP.
Why it matters: If you want to track specific directions in robot-agent exploration and planning without searching across the full program, this schedule consolidates titles, links, and session times so you can directly follow the relevant talks and posters (including late-breaking sessions and spotlights).
Signals: The list of late-breaking posters may grow as more are confirmed.
Source: http://bair.berkeley.edu/blog/2026/11/09/bair-at-corl-2026/